MDAug: Metamorphic Diffusion-based Augmentation for Improving Deep Learning-based Fault Localization Without Test Oracles

抄録

Deep Learning-based Fault Localization (DLFL) uses metamorphic testing to locate faults in the absence of test oracles. However, these approaches face the class imbalance problem, i.e., the violated data (i.e., minority class) is much less than the non-violated data (i.e., majority class). To address this issue, we propose MDAug: Metamorphic Diffusionbased Augmentation for improving DLFL without test oracles. MDAug combines metamorphic testing and diffusion model to generate the data of minority class and acquire class balanced data. We apply MDAug to three state-of-the-art DLFL baselines without test oracles, and the results show that MDAug significantly outperforms all the baselines in the absence of test oracles.